[ https://issues.apache.org/jira/browse/SPARK-19981?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15945166#comment-15945166
]
Mitesh edited comment on SPARK-19981 at 3/28/17 1:34 PM:
---------------------------------------------------------
As I mentioned on the PR, this seems like it should be handled here:
https://github.com/maropu/spark/blob/b5d1038edffff5d65a6ddec20ea6eef186d25fc3/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/Canonicalize.scala#L41
Perhaps it should handle canonicalizing an alias so {{a#1}} is the same as {{a#1 as newA#2}}.
Otherwise you have a similar problem with sorting. Here is a sort example with 1 partition.
I believe the extra sort on {{newA}} is unnecessary.
{code}
scala> val df1 = Seq((1, 2), (3, 4)).toDF("a", "b").coalesce(1).sortWithinPartitions("a")
df1: org.apache.spark.sql.Dataset[org.apache.spark.sql.Row] = [a: int, b: int]
scala> val df2 = df1.selectExpr("a as newA", "b")
df2: org.apache.spark.sql.DataFrame = [newA: int, b: int]
scala> println(df1.join(df2, df1("a") === df2("newA")).queryExecution.executedPlan)
*SortMergeJoin [args=[a#37225], [newA#37232], Inner][outPart=PartitioningCollection(1, )][outOrder=List(a#37225
ASC%NONNULL)][output=List(a#37225:int%NONNULL, b#37226:int%NONNULL, newA#37232:int%NONNULL,
b#37243:int%NONNULL)]
:- *Sort [args=[a#37225 ASC], false, 0][outPart=SinglePartition][outOrder=ArrayBuffer(a#37225
ASC%NONNULL)][output=List(a#37225:int%NONNULL, b#37226:int%NONNULL)]
: +- Coalesce [args=1][outPart=SinglePartition][outOrder=List()][output=List(a#37225:int%NONNULL,
b#37226:int%NONNULL)]
: +- LocalTableScan [args=[a#37225, b#37226]][outPart=UnknownPartitioning(0)][outOrder=List()][output=List(a#37225:int%NONNULL,
b#37226:int%NONNULL)]
+- *Sort [args=[newA#37232 ASC], false, 0][outPart=SinglePartition][outOrder=List(newA#37232
ASC%NONNULL)][output=ArrayBuffer(newA#37232:int%NONNULL, b#37243:int%NONNULL)]
+- *Project [args=[a#37242 AS newA#37232, b#37243]][outPart=SinglePartition][outOrder=ArrayBuffer(a#37242
ASC%NONNULL)][output=ArrayBuffer(newA#37232:int%NONNULL, b#37243:int%NONNULL)]
+- *Sort [args=[a#37242 ASC], false, 0][outPart=SinglePartition][outOrder=ArrayBuffer(a#37242
ASC%NONNULL)][output=List(a#37242:int%NONNULL, b#37243:int%NONNULL)]
+- Coalesce [args=1][outPart=SinglePartition][outOrder=List()][output=List(a#37242:int%NONNULL,
b#37243:int%NONNULL)]
+- LocalTableScan [args=[a#37242, b#37243]][outPart=UnknownPartitioning(0)][outOrder=List()][output=List(a#37242:int%NONNULL,
b#37243:int%NONNULL)]
{code}
was (Author: masterddt):
As I mentioned on the PR, this seems like it should be handled here:
https://github.com/maropu/spark/blob/b5d1038edffff5d65a6ddec20ea6eef186d25fc3/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/Canonicalize.scala#L41
Perhaps it should handle canonicalizing an alias so {{a#1}} is the same as {{a#1 as newA#2}}.
Otherwise you have a similar problem with sorting. Here is a sort example with 1 partition.
I believe the extra sort on {[newA}] is unnecessary.
{code}
scala> val df1 = Seq((1, 2), (3, 4)).toDF("a", "b").coalesce(1).sortWithinPartitions("a")
df1: org.apache.spark.sql.Dataset[org.apache.spark.sql.Row] = [a: int, b: int]
scala> val df2 = df1.selectExpr("a as newA", "b")
df2: org.apache.spark.sql.DataFrame = [newA: int, b: int]
scala> println(df1.join(df2, df1("a") === df2("newA")).queryExecution.executedPlan)
*SortMergeJoin [args=[a#37225], [newA#37232], Inner][outPart=PartitioningCollection(1, )][outOrder=List(a#37225
ASC%NONNULL)][output=List(a#37225:int%NONNULL, b#37226:int%NONNULL, newA#37232:int%NONNULL,
b#37243:int%NONNULL)]
:- *Sort [args=[a#37225 ASC], false, 0][outPart=SinglePartition][outOrder=ArrayBuffer(a#37225
ASC%NONNULL)][output=List(a#37225:int%NONNULL, b#37226:int%NONNULL)]
: +- Coalesce [args=1][outPart=SinglePartition][outOrder=List()][output=List(a#37225:int%NONNULL,
b#37226:int%NONNULL)]
: +- LocalTableScan [args=[a#37225, b#37226]][outPart=UnknownPartitioning(0)][outOrder=List()][output=List(a#37225:int%NONNULL,
b#37226:int%NONNULL)]
+- *Sort [args=[newA#37232 ASC], false, 0][outPart=SinglePartition][outOrder=List(newA#37232
ASC%NONNULL)][output=ArrayBuffer(newA#37232:int%NONNULL, b#37243:int%NONNULL)]
+- *Project [args=[a#37242 AS newA#37232, b#37243]][outPart=SinglePartition][outOrder=ArrayBuffer(a#37242
ASC%NONNULL)][output=ArrayBuffer(newA#37232:int%NONNULL, b#37243:int%NONNULL)]
+- *Sort [args=[a#37242 ASC], false, 0][outPart=SinglePartition][outOrder=ArrayBuffer(a#37242
ASC%NONNULL)][output=List(a#37242:int%NONNULL, b#37243:int%NONNULL)]
+- Coalesce [args=1][outPart=SinglePartition][outOrder=List()][output=List(a#37242:int%NONNULL,
b#37243:int%NONNULL)]
+- LocalTableScan [args=[a#37242, b#37243]][outPart=UnknownPartitioning(0)][outOrder=List()][output=List(a#37242:int%NONNULL,
b#37243:int%NONNULL)]
{code}
> Sort-Merge join inserts shuffles when joining dataframes with aliased columns
> -----------------------------------------------------------------------------
>
> Key: SPARK-19981
> URL: https://issues.apache.org/jira/browse/SPARK-19981
> Project: Spark
> Issue Type: Bug
> Components: SQL
> Affects Versions: 2.0.2
> Reporter: Allen George
>
> Performing a sort-merge join with two dataframes - each of which has the join column
aliased - causes Spark to insert an unnecessary shuffle.
> Consider the scala test code below, which should be equivalent to the following SQL.
> {code:SQL}
> SELECT * FROM
> (SELECT number AS aliased from df1) t1
> LEFT JOIN
> (SELECT number AS aliased from df2) t2
> ON t1.aliased = t2.aliased
> {code}
> {code:scala}
> private case class OneItem(number: Long)
> private case class TwoItem(number: Long, value: String)
> test("join with aliases should not trigger shuffle") {
> val df1 = sqlContext.createDataFrame(
> Seq(
> OneItem(0),
> OneItem(2),
> OneItem(4)
> )
> )
> val partitionedDf1 = df1.repartition(10, col("number"))
> partitionedDf1.createOrReplaceTempView("df1")
> partitionedDf1.cache() partitionedDf1.count()
>
> val df2 = sqlContext.createDataFrame(
> Seq(
> TwoItem(0, "zero"),
> TwoItem(2, "two"),
> TwoItem(4, "four")
> )
> )
> val partitionedDf2 = df2.repartition(10, col("number"))
> partitionedDf2.createOrReplaceTempView("df2")
> partitionedDf2.cache() partitionedDf2.count()
>
> val fromDf1 = sqlContext.sql("SELECT number from df1")
> val fromDf2 = sqlContext.sql("SELECT number from df2")
> val aliasedDf1 = fromDf1.select(col(fromDf1.columns.head) as "aliased")
> val aliasedDf2 = fromDf2.select(col(fromDf2.columns.head) as "aliased")
> aliasedDf1.join(aliasedDf2, Seq("aliased"), "left_outer") }
> {code}
> Both the SQL and the Scala code generate a query-plan where an extra exchange is inserted
before performing the sort-merge join. This exchange changes the partitioning from {{HashPartitioning("number",
10)}} for each frame being joined into {{HashPartitioning("aliased", 5)}}. I would have expected
that since it's a simple column aliasing, and both frames have exactly the same partitioning
that the initial frames.
> {noformat}
> *Project [args=[aliased#267L]][outPart=PartitioningCollection(5, hashpartitioning(aliased#267L,
5)%NONNULL,hashpartitioning(aliased#270L, 5)%NONNULL)][outOrder=List(aliased#267L ASC%NONNULL)][output=List(aliased#267:bigint%NONNULL)]
> +- *SortMergeJoin [args=[aliased#267L], [aliased#270L], Inner][outPart=PartitioningCollection(5,
hashpartitioning(aliased#267L, 5)%NONNULL,hashpartitioning(aliased#270L, 5)%NONNULL)][outOrder=List(aliased#267L
ASC%NONNULL)][output=ArrayBuffer(aliased#267:bigint%NONNULL, aliased#270:bigint%NONNULL)]
> :- *Sort [args=[aliased#267L ASC], false, 0][outPart=HashPartitioning(5, aliased#267:bigint%NONNULL)][outOrder=List(aliased#267L
ASC%NONNULL)][output=ArrayBuffer(aliased#267:bigint%NONNULL)]
> : +- Exchange [args=hashpartitioning(aliased#267L, 5)%NONNULL][outPart=HashPartitioning(5,
aliased#267:bigint%NONNULL)][outOrder=List()][output=ArrayBuffer(aliased#267:bigint%NONNULL)]
> : +- *Project [args=[number#198L AS aliased#267L]][outPart=HashPartitioning(10,
number#198:bigint%NONNULL)][outOrder=List()][output=ArrayBuffer(aliased#267:bigint%NONNULL)]
> : +- InMemoryTableScan [args=[number#198L]][outPart=HashPartitioning(10, number#198:bigint%NONNULL)][outOrder=List()][output=ArrayBuffer(number#198:bigint%NONNULL)]
> : : +- InMemoryRelation [number#198L], true, 10000, StorageLevel(disk,
memory, deserialized, 1 replicas), false[Statistics(24,false)][output=List(number#198:bigint%NONNULL)]
> : : : +- Exchange [args=hashpartitioning(number#198L, 10)%NONNULL][outPart=HashPartitioning(10,
number#198:bigint%NONNULL)][outOrder=List()][output=List(number#198:bigint%NONNULL)]
> : : : +- LocalTableScan [args=[number#198L]][outPart=UnknownPartitioning(0)][outOrder=List()][output=List(number#198:bigint%NONNULL)]
> +- *Sort [args=[aliased#270L ASC], false, 0][outPart=HashPartitioning(5, aliased#270:bigint%NONNULL)][outOrder=List(aliased#270L
ASC%NONNULL)][output=ArrayBuffer(aliased#270:bigint%NONNULL)]
> +- Exchange [args=hashpartitioning(aliased#270L, 5)%NONNULL][outPart=HashPartitioning(5,
aliased#270:bigint%NONNULL)][outOrder=List()][output=ArrayBuffer(aliased#270:bigint%NONNULL)]
> +- *Project [args=[number#223L AS aliased#270L]][outPart=HashPartitioning(10,
number#223:bigint%NONNULL)][outOrder=List()][output=ArrayBuffer(aliased#270:bigint%NONNULL)]
> +- InMemoryTableScan [args=[number#223L]][outPart=HashPartitioning(10, number#223:bigint%NONNULL)][outOrder=List()][output=ArrayBuffer(number#223:bigint%NONNULL)]
> : +- InMemoryRelation [number#223L, value#224], true, 10000, StorageLevel(disk,
memory, deserialized, 1 replicas), false[Statistics(47,false)][output=List(number#223:bigint%NONNULL,
value#224:string%NULL)]
> : : +- Exchange [args=hashpartitioning(number#223L, 10)%NONNULL][outPart=HashPartitioning(10,
number#223:bigint%NONNULL)][outOrder=List()][output=List(number#223:bigint%NONNULL, value#224:string%NULL)]
> : : +- LocalTableScan [args=[number#223L, value#224]][outPart=UnknownPartitioning(0)][outOrder=List()][output=List(number#223:bigint%NONNULL,
value#224:string%NULL)]
> {noformat}
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org
|